In briefAdaptive curriculum and guarded anti-curriculum training for dense prediction
SummaryThe paper proposes PaceShift, a two-phase training method for dense prediction. The first phase selects and weights samples by estimated supervision reliability, augmentation consistency, model disagreement, and prediction certainty to stabilize learning. The second phase ranks predictions from multiple spectral views and refines them with a pairwise preference objective against a frozen reference model. The authors report accuracy gains, improved calibration and corruption robustness in the evaluated settings, with no added inference computation.
In briefReducing shortcut reliance in chain-of-thought reasoning with contrastive decoding
SummaryThis paper studies whether large language models rely on query-specific shortcuts instead of their generated reasoning steps during chain-of-thought prompting. It applies context-aware decoding as Rationale Contrastive Decoding, which contrasts next-token distributions with and without the rationale and amplifies tokens whose likelihood rises due to it. Across five reasoning benchmarks, the method improves accuracy over standard chain-of-thought, and intervention analyses show final answers depend more on rationale content.
In briefAutoPhyX: Automatically generating text-described physical properties for 3D objects
SummaryAutoPhyX is a framework that predicts spatially varying physical properties for 3D objects from text descriptions, aiming to support more accurate 4D physical simulation. The authors built a data pipeline that splits 3D assets into semantic parts and pairs each with physically plausible parameters and text. The model uses cross-modal modulation, where text features adjust visual features, and works in a voxel field compatible with meshes, point clouds, Gaussian Splatting, and NeRF. The paper reports high accuracy and generalization to in-the-wild objects in a single forward pass.
In briefBRIDGE: A Gate That Blocks False Improvements in Model Updates
SummaryThe paper proposes BRIDGE, a model-agnostic acceptance gate that decides whether to accept each model update using predictions and available labels, reducing selection bias through cross-fitting and requiring the gain's lower bootstrap bound to exceed a margin. Across multiple datasets and retrospective clinical tasks, it yields far fewer false improvements than the naive rule of accepting every validation gain. Mean gain per deployed update is also higher than under the naive rule.
In briefA unified Stein-operator framework for estimating unnormalized models
SummaryThe paper builds a unified framework linking proper scoring rules and Stein's method for estimating unnormalized statistical models. It constructs a class of proper scoring rules that includes score matching, ratio matching, pseudolikelihood, and noise contrastive estimation. The authors show these scoring rules naturally induce Stein operators, and use the associated geometry to give a unified characterization of the asymptotic efficiency of the resulting estimators.
In briefShortcut-Resistant CAM Distillation for Long-Tailed Recognition
SummaryLong-tailed datasets make it hard for models to generalize to rare classes, which tend to rely on background or texture shortcuts. The paper proposes SRCD, a plug-and-play framework that distills object-focused class activation map (CAM) explanations from head classes to tail classes. It aggregates head-class CAMs into a teacher using energy-based weighting and transfers it to the tail-class CAM. Experiments on long-tailed benchmarks consistently improve strong baselines.
In briefKeeping Time vs. Understanding Time: Why Neither Explains LLM Temporal Reasoning
SummaryThis paper challenges the common split of temporal reasoning into arithmetic and semantic modes. Finetuning on benchmark subsets mostly yields dataset-specific gains, with little transfer to other datasets in the same mode. Cognitive-demand evaluation finds that knowledge, rather than quantitative or logical reasoning, is the main driver of performance. The authors conclude that current benchmarks largely reward models for remembering how time typically works.
Auss Abbood, Fernando Mart\'inez-Plumed, Jos\'e Hern\'andez-Orallo et al.
In briefAsynchronous, Generation-Free ABC: Streaming AMIS Reweighting for Heterogeneous HPC Simulations
SummaryThis paper proposes a generation-free approximate Bayesian computation (ABC) sampler that updates its proposal as each simulation finishes and dispatches new candidates to idle workers, avoiding synchronization barriers. It reuses the full sample history and reweights particles against a reconstructed mixture of proposals to estimate the posterior. The method is compared with a pyABC baseline on a Cellular Potts tissue simulator, reporting throughput and posterior results.
In briefRethinking how AI model orchestration improves time series forecasting
SummaryTime series foundation models perform unevenly across forecasting tasks, and agentic model orchestration (AMO) selects or combines them. This paper introduces FairAMO, a controlled protocol with frozen candidate forecasts, to isolate the effect of orchestration itself. On GIFT-Eval, AMO's accuracy gains are bounded by the candidate models, with the best orchestrators reducing MASE by less than 6% versus the best single model. Gains depend on complementarity among candidates rather than their number, and a few recent historical windows capture all of the benefit.
In briefDiverse Parallel Thinking (DPT): Reinforcement learning that keeps language model reasoning paths varied
SummaryThis paper addresses a problem in which language models trained with reinforcement learning for parallel thinking produce reasoning paths that become similar to one another. The authors propose Diverse Parallel Thinking (DPT), which penalizes the prefix similarity of parallel reasoning paths to encourage distinct reasoning strategies. Experiments on multiple reasoning benchmarks reportedly show that DPT outperforms previous parallel thinking methods, and the authors identify path diversity as a key factor.
In briefA Strong MoE Is All You Need: Using Expert Routing as Process Supervision for Reasoning
SummaryProcess reward models need reliable step-level labels for multi-step reasoning, but obtaining them is costly. The paper shows that routing statistics from Mixture-of-Experts models act as a process-confidence signal competitive with output-logit confidence. It proposes RoCo-Rethink, which regenerates the trajectory after the least confident step, and RoCo-PRM, which uses a strong MoE to auto-annotate steps and distill them into a standalone PRM. Both reportedly outperform strong baselines across several MoE architectures and reasoning tasks.
In briefBalancing foundation models and specialized learning for forehead-crease biometric recognition
SummaryFacial recognition weakens when parts of the face are hidden, but the forehead usually stays visible and has distinctive crease patterns that could serve as a supplementary biometric. The paper evaluates zero-shot reasoning with multimodal large language models (FIRE), domain-adapted learning, and trait-specific modeling for forehead-crease recognition. The best zero-shot model, GPT-4o+FIRE, reaches 18.22% EER, while Edge-Forehead reaches 8.30% EER and 54.01% TMR at 0.1% FMR. The authors conclude that trait-specific learning is more effective than MLLM-based zero-shot verification for robust forehead-crease recognition.
In briefAn exact 93-move solution to a puzzle on an 11×11 crystal-exchange board
SummaryThe paper models one level of the game Order of the Sinking Star as a crystal-exchange puzzle on an 11×11 board and determines the minimum number of directional commands needed to solve it. The authors give an explicit 93-command solution and use an exact, cost-bounded breadth-first search to exclude any shorter solution. The result is a computer-assisted proof for this specific instance, with source code, logs, and a reproduction driver provided.
In briefRubricOpt: Letting a Model Revise Its Own Training Criteria
SummaryThis paper introduces RubricOpt, a framework in which a language model diagnoses its own failures and revises the behavioral rubric guiding its preference optimization. Training uses Direct Preference Optimization (DPO) with preference data and objective held fixed, and the rubric is removed at evaluation. Experiments on capability and open-ended generation benchmarks show that rubric conditioning can improve over ordinary DPO, with gains and regressions varying across models and datasets.
In briefTenetBench: A Russian Corpus Testing Whether Language Models Follow Professional Ethics Codes
SummaryThe paper introduces TenetBench, a Russian-language corpus of 3,433 requests that conflict with clauses of fifteen professional codes of ethics, each paired with a benign twin differing only in legitimacy. An automated pipeline generates the tasks from code text, and human specialists largely judged the conflicting tasks plausible. Fourteen model and reasoning-mode combinations were tested on whether they produce conflicting content, decline, or warn, and eleven of the fourteen produced conflicting content on most such requests.
WhyProposes a clause-anchored Russian benchmark with automated generation, and evaluates fourteen model configurations with human-validated labels, giving solid evidence.
Yegor Anichkov, Valery Vinogradov, Andrey Furov et al.
In briefGenerative modeling using nonlinear Schrödinger bridges and branching diffusions
SummaryThis paper proposes a generative modeling framework based on forward-backward PDEs related to mean-field game theory and nonlinear Schrödinger systems. The authors derive SBNL, a nonlinear extension of the classical Schrödinger bridge problem, via a coupled PDE system. They design a parallel, high-dimensional simulation algorithm using branching diffusions and probabilistic representations of semilinear PDEs. Experiments cover recovering the classical bridge, nonlinear extensions, and image generation on MNIST and FFHQ.
In briefDensity-level backward error for allocating compute in low-step generative flow sampling
SummaryThis paper studies how to distribute a fixed budget of network evaluations along a generative ODE flow when sampling with very few steps. The authors propose a density-level backward-error framework that keeps only the solver error affecting the evolving marginal distribution, yielding a solver-order-aware compute-allocation rule. On a pretrained CIFAR-10 EDM probability-flow ODE, the density-aware allocation improved fixed-update Euler sampling in the low-NFE regime.
In briefExtending Graspable-Part Segmentation for Wearable Robots to Multi-Object Scenes
SummaryThis paper refines existing vision-based portable systems that recognize graspable object parts to help upper-limb prostheses select grasps. It adds a laser pointer to the setup to disambiguate target objects and retrains the deep convolutional neural networks, using mixed reality to generate realistic images. Experiments report up to 9% better target disambiguation accuracy than existing solutions, with computational cost preserved.
Simone Lugani, Matteo Montani, Paolo Gastaldo et al.
In briefAsymmetry Tunneling: Capturing Symmetry-Breaking Structure Without Adaptive Equivariance Control
SummaryEquivariant neural networks build in prescribed symmetries, but real data may break those symmetries. The authors add a residual path beside a strictly equivariant model and restrict how the two connect, to study how task-relevant structure passes through. Across five benchmarks, the residual path sometimes collapses and sometimes survives, a survival the authors call asymmetry tunneling, and they characterize it with four diagnostics. They conclude that adaptively controlling equivariance relaxation is not required to capture potential symmetry-breaking structure.
In briefOmniSimulator: Training small language models to simulate real users' heterogeneous behavior
SummaryThis paper presents OmniSimulator, a training framework that helps small LLMs predict realistic user behavior on short-video platforms. It uses a five-dimension structured reasoning schema and a teacher model to generate high-quality reasoning traces, so small models learn behavior-relevant patterns instead of superficial correlations. The authors built 9,881 training instances from three months of interaction data from 200 real users. Experiments report that Qwen3-8B's average score improved by 34.33%, competitive with Claude-Sonnet-4.5, while supporting smaller self-hosted deployment.}
In briefPreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents
SummaryThis paper proposes PreCoMem, a memory consolidation framework inspired by predictive coding that stores a dialogue agent's knowledge of the user as confidence-weighted beliefs. It uses an Effective Surprisal signal, combining semantic deviation and contradiction evidence, to route each turn to one of three updates: maintain, profile as a low-confidence hypothesis, or correct contradicted facts. The authors report consistent state-of-the-art accuracy on LoCoMo, LongMemEval-S, and PersonaMem-v2, and that their new DiMoBench shows adaptation to belief drift without overreacting to noise.
In briefRethinking video storage for AI training: breaking cross-frame dependencies for fast random access
SummaryAI training often samples random frames from large video datasets, but standard video formats are built for sequential playback and rely on complex inter-frame references that slow decoding. This paper proposes dependency-aware sparse decoding, which reconstructs only requested frames and their required references, plus an optional offline encoding called DDRA that limits each interior frame to two intra-coded anchors. On a large robotics dataset, at comparable reconstruction quality, the method achieves 1.45 times the loading throughput of LeRobot v3 with 58.7% less media payload.
In briefHarness Scaling: training agents across varied execution frameworks to generalize to unseen ones.
SummaryThis paper studies Harness Scaling, a way to scale agentic reinforcement learning by structuring variation in the agent harnesses (execution frameworks for tools, context, and interaction) that policies run within. The authors build a factorized harness space, pick a compact training subset with balanced factorial coverage, and assign tasks across harnesses using locality-aware organization. Across multiple benchmarks, model scales, and RL algorithms, the approach improves generalization to unseen harness combinations, and policies trained with structured white-box variation transfer to unseen external black-box harnesses.
In briefRegoMender: A Multi-agent Framework for Automatically Repairing Rego Policies
SummaryThis paper addresses automated repair of Rego policies used in Policy-as-Code for cloud-native systems. The authors introduce RegoBench, a benchmark of 1,159 real-world defects, and RegoMender, a feedback-guided multi-agent framework that collaborates through a shared blackboard and atomic toolkit. The framework outperforms the strongest baseline on RegoBench and reaches an 86.4% repair rate on 132 real-world CVE vulnerabilities.
In briefOmnilingual-MsASR: A Benchmark for Multilingual, Multimodal, Long-Form Multi-Speaker Speech Recognition in Omni LLMs
SummaryThis paper introduces Omnilingual-MsASR, a benchmark for testing omni-modal large language models on long-form, multi-speaker speech recognition. It covers six languages, recordings up to 30 minutes long, and scenes with up to 20 speakers, providing parallel audio and video inputs. Evaluations of leading models show large performance gaps, with tcpWERs often exceeding 100%, highlighting temporal localization, speaker disambiguation, and cross-modal synergy as key challenges.
In briefBlazeVGGT: Training-free, camera-token-based sparsification makes VGGT about 13× faster at 2,000 frames
SummaryVGGT is a feed-forward visual geometry model that reconstructs multiple views through dense global attention, but its cost grows quadratically with the number of input views. The authors observe that shallow camera tokens capture local view neighborhoods while deeper tokens encode global inter-view geometry. Based on this, they propose BlazeVGGT, which uses greedy determinantal point process selection on camera-token similarity to choose complementary views as key-value context, while keeping all frames as queries, and can be plugged into pretrained VGGT without parameter updates. The paper reports about 13× faster inference at 2,000 frames and a Chamfer Distance reduction from 0.41 to 0.32, along with compatibility with FastVGGT and LiteVGGT.
WhyTraining-free key-value view selection via camera-token similarity and determinantal point processes cuts cross-view attention; validated on reconstruction, pose, and several VGGT variants.
In briefDynaWeave: Training AI Agents to Adapt to Changing Situations
SummaryThis paper introduces DynaWeave, a framework that automatically turns existing static tasks into verifiable dynamic scenarios with external events such as changed requirements, environment states, or new information. Event dependencies are modeled with directed acyclic graphs, and a verifier updates only the checks affected by events that actually occur. The authors collected 2,095 verified interaction trajectories across 36 environments and used them to fine-tune Qwen3.5 models. Experiments on Gaia2, CostBench, and ToolMaze report consistent overall improvements across three model sizes.
In briefBeyond Shortcut Learning: Hierarchical Spatial Prototype Regularization for Face Forgery Detection
SummaryThis paper addresses the poor generalization of deepfake detectors to unseen manipulation methods. It reports that naively fine-tuned vision models suffer forensic representation collapse, where extracted cues become entangled with unstable, dataset-specific artifacts. The authors propose Hierarchical Spatial Prototype (HSP) regularization, which learns a structured prototype subspace over spatial features, along with prototype orthogonality and second-order assignment balance. The paper reports improved frame-level AUC in cross-dataset and cross-manipulation settings, and gains when HSP is used as a plug-in module.
In briefAllocating attention sparsity by output sensitivity for video diffusion transformers without training
SummaryFull 3D attention is expensive in video Diffusion Transformers, and block-sparse attention reduces cost by computing only selected blocks. This paper proposes ALLOT, a training-free method that allocates sparsity across layer-step pairs by estimating each pair's contribution to final-video error through cell probing and trajectory calibration, then optimizing densities under a measured-latency budget. At equal or shorter wall-clock time, ALLOT outperforms the compared allocation strategies with XAttention, gaining 11.8 dB PSNR on HunyuanVideo. It also outperforms or ties the default in all nine settings across Wan2.1, Wan2.2, and HunyuanVideo.
In briefHow far can large foundation models be shrunk into small, fast jet taggers?
SummaryThis paper studies jet tagging at the Large Hadron Collider by using knowledge distillation to transfer six teacher models, including two foundation models, into five small student models. Every teacher improves every student, and the size of the gain depends mainly on teacher strength. A distilled L-GATr-slim student with 20 thousand parameters outperforms every published jet tagger trained from scratch, and with two thousand parameters it matches ParT, which has a thousand times more parameters.
WhySystematically distils six teachers into small students, with a 2k-parameter student matching a model a thousand times larger, backed by thorough experiments.
In briefEvents, Not Pixels: Where World Models Get Returns Wrong
SummaryWorld models used for planning are usually judged by how well they render frames, yet visually faithful rollouts can still assign wrong returns. The authors introduce a transition-level diagnostic that swaps imagined transitions for real ones and compares the return-error reduction against a budget-matched random control. They find that reward and episode-ending event transitions are rendered about as well as other steps, but restoring them removes a median of 70% of return error, while restoring transitions with the largest pixel-level error does worse than random.
WhyShows return error concentrates at rare reward events via transition-level diagnostics; statistically significant across 44 model-collection pairs, offering a new evaluation lens for planning.
In briefTesting whether vision-language models can point to the right image region when helping visually impaired users
SummaryThis paper studies visual grounding in vision-language models (VLMs) for visually impaired people, meaning whether a model can identify the image region relevant to a question. The authors evaluate open-source VLMs on the VizWiz-VQA-Grounding dataset and challenge, which uses real photos taken by visually impaired users. They also propose VG-SHAP, an extension of an existing explainability framework that lets black-box VLMs be evaluated for grounding. The results show that models often give correct answers but fail to ground them in the proper image region, though the authors note this is limited by the segmentation model used in their framework.
In briefBeyond the Question: A Preserve-Focus-Fuse Method for Slide Image Q&A
SummaryThis paper introduces BEYONDPATH, a framework for answering questions about whole-slide pathology images. It preserves question-agnostic slide morphology, retrieves question-specific evidence from foreground patches, and fuses both into a fixed 96-token visual prefix. The authors report an Avg-9 score of 0.769 on WSI-Bench and 0.639 zero-shot overall accuracy on SlideBench-BCNB. Ablations indicate that morphology preservation and question-directed retrieval each contribute benefits.
In briefMERP: A Latent-Bottleneck Model for Robust Multimodal Emotion Recognition
SummaryThis paper proposes MERP, a model that maps text, visual, and acoustic inputs into a fixed-size latent representation to recognize multiple emotions. When 20 percent of modalities are randomly dropped at test time, MERP reaches micro-F1 scores of 0.563 on CMU-MOSEI and 0.780 on M3ED, above the strongest evaluated baselines. In the audio-only CMU-MOSEI setting, it keeps 0.473 micro-F1, compared with 0.022 for LDDU.
In briefOne Act, Many Masks: Policy optimization for LLM agents that credits equivalent actions consistently
SummaryThis paper studies reinforcement learning for LLM agents on long multi-turn tasks, noting that different text responses can produce the same executable effect while credit remains tied to each wording. The authors call this credit-update misalignment and propose Effect-Marginal Policy Optimization (EMPO), which couples updates across alternative textual realizations through an effect-level surrogate. The method works with existing trajectory-level and step-level credit estimators without extra critics or rollouts. Experiments on ALFWorld, WebShop, and Sokoban report better task performance and learning efficiency than strong baselines.
In briefBIRD-RL: Scaling Agentic RL for Stateful Data Environments
SummaryThis paper introduces BIRD-RL, a four-stage agentic reinforcement learning curriculum for data-centric tasks such as SQL debugging, where environments are stateful. It adds a Trajectory-scoped Persistent Execution Infrastructure that allows database replicas to be reused while keeping each trajectory isolated, reported as a 12.8x gain in spatial efficiency. On the BIRD-CRITIC-SQLite benchmark, the 7B and 14B models reach 44.40% and 48.00% success rates, compared with 48.20% for a Claude-Opus-4.6 agent.
In briefGeneralization bounds from a vector-valued L_p correlation measure V_α
SummaryThe paper studies generalization of randomized learning algorithms using a one-parameter family of dependence measures, V_α, the vector-valued L_p correlation. The resulting bounds interpolate between total variation at α=1 and variance/χ²-type behavior at α=2. Experiments on Bayesian linear regression, noisy MNIST, and a pretrained ResNet-50 on CIFAR-10 show how tuning α changes bound tightness.[]
Mohammad Hossein Momeni Hamaneh, Mohammad Mahdi Mojahedian, Mohammad Hossein Yassaee
In briefKAN-Based Nonlinear Factorization for Compressing Embedding Layers
SummaryThis paper proposes improving low-rank factorization of embedding matrices by adding learnable nonlinear projections to the compressed embedding layer, inspired by Kolmogorov-Arnold Networks (KAN). The method is integrated into the Skip-gram framework and compared against a standard low-rank baseline under matched parameter budgets. Evaluation uses Word Analogy and Semantic Similarity benchmarks, and the authors also introduce an initialization strategy from a pre-trained uncompressed embedding layer. The results indicate that nonlinear projections yield more faithful compressed embeddings at fixed dimensionality.
In briefWhen Thoughts Become Memory: Unifying In-Context Learning and Fine-Tuning
SummaryThe paper gives a certified rule for deciding whether unseen data should go into a pretrained model's context or its weights. The rule rests on an exact law for the information value of that data, and an anytime-valid test uses held-out queries to make the choice with controlled error. The authors test it across several models and tasks and analyze when learning in context helps and when it does not.
WhyProposes a certified rule, built on an exact information-value law, for choosing context versus weights; novel, with broad validation across models and tasks.
In briefAlternative Active Learning Strategies with a Bayesian Bradley-Terry Model
SummaryThis paper studies how to recover rankings from a limited number of noisy pairwise comparisons. The authors propose a Bayesian Bradley-Terry model (BBTM) that uses Polya-Gamma data augmentation for Gibbs sampling and quantifies uncertainty in abilities and ranks. Three active learning strategies based on the knowledge gradient principle are introduced, along with Posterior Separation Reliability (PSR), an internal ranking-reliability indicator that needs no reference ranking. In synthetic experiments, BBTM generally recovers rankings more accurately than preference-based Bayesian comparative judgement.
In briefCCG Diffusion: Syntax-guided text-to-image generation using Combinatory Categorial Grammar
SummaryText-to-image models often misplace attributes, spatial relations, and counts because they do not receive the prompt's syntactic structure. This paper proposes CCG Diffusion, a training-free method that uses Combinatory Categorial Grammar to read modification relations and resolve prepositional phrase attachment ambiguity. The method applies prompt rewriting and attention guidance on Stable Diffusion and Sana. The authors report the highest scores among compared methods for attribute binding and 2D spatial relations, and significant gains for prepositional phrase ambiguity resolution.
Ryosuke Yamaki, Daichi Mochihashi, Nobutaka Shimada et al.
In briefSolving Go Under Positional Superko: Sound Proof Transfer and New Board Values
SummaryThis paper studies Go under positional superko, a rule forbidding any repeated whole-board position, which earlier solved boards mostly did not use. The authors present Sanko, a parallel proof-number solver that soundly keeps transposition tables, life-and-death recognition, and move pruning under this rule. It reports solving 5x5 Go in 16 minutes on six threads and computing new values for twenty-three boards, eleven previously unsolved under any rules. The authors also release their solver source and run records for reproducibility.
WhyIntroduces a sound proof-transfer theorem resolving PSK solving difficulties, with new values for unsolved boards and substantial experimental support.
In briefBeyond Small Talk: How Disclosure History Shapes AI Conversation
SummaryThis paper introduces TraDE, an evaluation framework based on Social Penetration Theory that uses predefined disclosure trajectories to test how earlier disclosure history affects a model's replies to the same current disclosure. Across several models and benchmarks, responses differed by preceding trajectory, especially in whether models accept or challenge the user's framing. These differences track accumulated exposure to deeper disclosure and persist for several turns after disclosure returns to shallow levels.
In briefA Gap Between Prediction and Optimization in GuanDan Self-Play
SummaryThe paper asks whether outcome-predictive signals are reliable targets for reinforcement learning. In the multi-round imperfect-information card game GuanDan, it uses changes in a frozen match-win predictor as round-level supervision within DanZero-style Monte Carlo self-play. One selected checkpoint clearly outperforms its baseline, but none of five replication seeds is positive and the pooled ten-seed interval includes zero. The authors conclude that such a signal can describe outcome-relevant progress without reliably providing a beneficial optimization direction.
In briefECHO: Hidden shared offsets in preference optimization cause vocabulary collapse
SummaryThe paper finds that preference optimization can raise ranking metrics while building a shared directional shift in hidden states that the relative ranking loss largely ignores. This common-mode offset projects onto the output layer, elevating one token's probability until it dominates decoding and collapses generation. The authors propose this representation-level signal as a way to warn of generation collapse before it occurs.
WhyIdentifies a common-mode hidden offset invisible to ranking loss causing vocabulary collapse, with injection and masking experiments providing causal evidence across many runs.
In briefLLMSeg: Training-free text-prompted image segmentation using B-splines
SummaryThe paper introduces LLMSeg, a method that turns a general vision-capable language model into a text-prompted image segmenter without extra segmentation heads or training. It works in three stages: detecting instances matching a text query, tracing each one with closed periodic B-splines, and iteratively refining errors. Evaluated on referring and reasoning segmentation benchmarks, LLMSeg is reported to be competitive with the specialist SAM 3 model and reaches 71.5 cIoU on ReasonSeg validation.</summary>
WhyRepurposes a general MLLM for text-prompted segmentation via a training-free B-spline pipeline, matching or beating SAM 3 across many benchmarks.
In briefLiveAct: Real-Time Multi-Agent World Models with Mixed Control Types
SummaryThis paper studies video world models that generate a shared visual future for several controlled agents, each receiving different kinds of controls, from game buttons to robot commands. It proposes LiveAct, built around a Heterogeneous Action Interface that binds each action stream to its intended visual agent and uses language plus synchronized values to interpret each control variable. The authors report agent-attribution accuracy of 0.898 for games and 0.942 for robot manipulation, and a causal generator running at 17.70 FPS at 512×448 resolution.
In briefVideoGen Model 1: Comparing Prompt Expansion and Formatting for Audio-Video Generation
SummaryThis paper reports a small controlled pilot of VideoGen Model 1, a post-trained version of MiniMax H3, testing 30 prompts in raw, formatted, and creatively expanded forms. The generator, seed, and sampling settings were held fixed, and each clip was rated on five dimensions whose mean gave an overall score. Compared with formatting alone, expansion raised overall scores by 0.63 points under VA-Judger, while the 0.15-point gain under Gemini 3.8 Flash had an interval including zero and remained uncertain.
David Grossman (VideoGen), Anton Koenig (VideoGen)
In briefMulti-view inspection combining deep learning with rule-based decision fusion for component acceptance
SummaryThis paper proposes a region-aware optical inspection framework that pairs instance segmentation of seat-region defects with image classification of shaft-region anomalies. A deterministic decision fusion layer, built on formalized industrial standard operating procedures, converts model outputs into component-level decisions. Evaluated on 380 automotive fuel injector needles using 360-degree multi-view imaging, the approach reduced the false rejection rate from 57.6% under simple union logic to 10.9%, while keeping non-conforming recall at 100%.
Srivishnu Ethirajulu Krishnaraj, Rishikesh R. Chopade
In briefLayer-wise neuron alignment via temperature-controlled spectral sharpening for weight-space matching
SummaryThis paper addresses how permutation symmetry makes independently trained, functionally equivalent neural networks appear distant in weight space, causing large loss barriers under interpolation. The authors propose a progressive layer-wise method that aligns a correlation matrix to a Hungarian-algorithm anchor and sharpens it toward a permutation matrix via temperature-controlled spectral annealing, followed by weight matching refinement. Experiments on MLPs, CNNs, and NeRFs report reliable neuron correspondence recovery and reduced loss barriers for NeRF weight interpolation compared with existing approaches.
In briefDetection Alone Is Not Enough: Checking Reasoning Traces as Testable Program Hypotheses
SummaryThe paper asks how far a language model's single reasoning trace can be verified without a reference solution, judge model, or weight updates. On a frozen LLaMA-2-7B-chat evaluated on GSM8K, self-checking reaches an AUROC of 0.741, and most wrong traces contain no false equation. The authors propose a metamorphic probe that turns each trace into a program and tests it under perturbation, but no tested policy significantly improves accuracy. They conclude that repairing wrong traces, not detecting them, is the open problem.
In briefNominate-then-Adjudicate: LLM-guided one-pass tuning of MIP solver settings
SummaryThis paper addresses the problem that mixed-integer programming solvers have many parameters whose best values vary by instance, and per-instance search is costly. It proposes a pipeline that samples a fixed portfolio of configurations once, then uses a foundation model to select one configuration per instance in a single call, with no solver calls at inference. On the evaluated problem families, it wins more often than it loses against the top nominee and the default, and outperforms per-instance baselines such as ISAC+ under a matched one-pass protocol.
In briefFM-Occ: Feature-Diverse Upsampling and Memory-Aware Decoding for 3D Occupancy Prediction
SummaryThis paper addresses 3D semantic occupancy prediction, which estimates occupancy and semantic labels for each voxel in a scene. The authors found that conventional sparse transposed-convolution upsampling produces poorly differentiated voxel responses, and that later decoder layers discard earlier correct semantic predictions. They propose FM-Occ, combining 3D Sparse Subvoxel Shuffle (3S) and Entropy-Regulated Layer-wise Memory (ER-LM), which raises mIoU on SemanticKITTI from 12.16% to 12.73%.
Yujia Zhai (1), Guoming Huang (1), Zhuanghuai Wang (1) et al.
In briefA parallelizable probabilistic latent reasoning method for generating diverse reasoning paths
SummaryLanguage models that use verbalized chain-of-thought generate text token by token, and many tokens serve fluency rather than reasoning. This paper proposes Probabilistic Latent Reasoning Models (PLRM), a parallelizable variational framework that models a distribution over multiple reasoning trajectories in a continuous latent space, using a discrete strategy variable and a mixture-based prior. The authors report that PLRM performs competitively with deterministic latent reasoning models on single-trajectory datasets, outperforms stochastic baselines on multi-trajectory datasets, and trains more efficiently.
In briefCascading Bandits with Noisy Observations: Algorithms and Regret Bounds (a learning method for choosing arms in sequence under noisy feedback and costs)
SummaryThe paper proposes a stochastic cascading bandit model in which probing an ordered shortlist of K out of L arms carries position-dependent costs, and the learner sees only noisy feedback. For noise models where per-arm scores allow the optimal shortlist to be found by sorting, a UCB-based algorithm achieves O((L-K)log T) instance-dependent regret, matching an algorithm-agnostic lower bound up to constant factors. When such scores do not exist, the authors introduce DP-UCB, which embeds optimism in an exact dynamic program and attains O(LK log T) regret. Theoretical results are validated through numerical experiments on synthetic and real-world datasets.
In briefAdvFidelity: Combining Human-Preference Alignment and Fast Image Generation via Reinforcement Learning
SummaryThis paper proposes AdvFidelity, which adds an artifact-sensitive adversarial reward to reinforcement learning so that image generators can align with human preferences and produce high-fidelity images in few sampling steps within a single training stage. It avoids distribution matching distillation and other explicit distillation objectives. A discriminator learns to rank high-quality references above reconstructions made by the current generator, which highlights low-level artifacts. Using 4 or 8 steps without CFG, the method matches full-step LeapAlign in alignment and overall generation quality.
In briefRight On Q: Recovering low-bit LLM reasoning via self-generated trajectories
SummaryThe paper proposes OnQ, a pipeline that recovers reasoning ability in low-bit quantized large language models. The quantized student generates its own responses and receives full-precision teacher supervision on those prefixes, using a candidate-wise Dynamic KL that weights forward and reverse KL by teacher-student probability ratios. Across Qwen3-1.7B, 4B, and 8B at 2- and 3-bit precision, the authors report the highest nine-benchmark average, with the largest gains on long-horizon reasoning. At 2 bits on Qwen3-8B, the paper reports a 10.26-point average improvement over the strongest baseline, along with shorter responses and fewer truncations.
In briefAOV-GS: Active open-vocabulary mapping of 3D scenes with Gaussian splatting
SummaryThis paper presents AOV-GS, a robot mapping system that uses frozen SAM and CLIP models to guide exploration and build 3D Gaussian maps supporting open-vocabulary text queries without manual masks. Compared with geometry-only active planning in matched controlled tests, it raises average language-field mIoU by 1.3 points on Replica and 1.8 points on Matterport3D. On Replica, depth L1 also improves, while novel-view PSNR stays comparable.
In briefCombining X-ray and neutron scattering data with modular neural networks for structure inference
SummaryThis paper proposes a modular probabilistic framework that combines multiple scattering measurements to infer structure. Separate posterior estimators are trained for each observation type, such as X-ray, time-of-flight neutron, and polarized neutron reflectometry, and are fused at inference time with a compositional sampler, without joint retraining. The approach is applied to depth-resolved nanometer-scale analysis of thin films and interfaces, including homogeneous composition across neutron contrasts and heterogeneous composition of X-ray and neutron data.
Valentin Munteanu, Alexander Hinderhofer, Dmitry Lapkin et al.
In briefSeeing Through Denoising: Using Background Reconstruction to Detect Tiny Infrared Targets
SummaryThis paper introduces DiffIRSTD, a generative framework for detecting small targets in infrared images. It treats targets as noise-like signals and uses a diffusion model to reconstruct a clean, target-free background, then highlights targets using the residual against the original image. A one-step denoising strategy and a frequency-guided perception loss improve efficiency and reconstruction quality. The authors report state-of-the-art results on IRSTD-1K, NUAA-SIRST, and NUDT-SIRST.
In briefReading the Mask: BI-RADS-Based Rationale Generation for Breast MRI Segmentation
SummaryThis paper derives clinically relevant BI-RADS descriptors from breast MRI segmentation masks and dynamic contrast-enhanced series, using a decision tree with eight radiomics measurements plus lesion size and component count. A large language model then generates natural-language rationales restricted to these derived records. Applied to 1,506 MAMA-MIA cases, it produced 3,012 rationales with 99.98% faithfulness and 99.3% counterfactual accuracy. The rationales are intended as drafts for radiologist review, not autonomous reports.
Hrishikesh Singh, Aarti Sharma, Divyanshu Kumar et al.
In briefMemory-Centric Continual Learning (MCCL) for AI agents
SummaryThe paper argues that current external memory systems for AI agents act mainly as passive retrieval caches and do not clearly enable continual learning. It proposes Memory-Centric Continual Learning (MCCL), which judges memory by whether it improves future actions. MCCL includes a three-axis taxonomy (substrates, assets, policies), a two-loop lifecycle, an evaluation protocol for retention, adaptation and transfer, and a memory scaling framework.
In briefGaP-MoE: Making Mixture-of-Experts Experts Specialize by Projecting Their Gradients
SummarySparse Mixture-of-Experts models route each input to a few experts, but these experts can become similar and lose diversity. GaP-MoE tackles this by changing expert gradients instead of the loss: each step picks a leader expert and projects the other experts' gradients orthogonal to the leader's recent gradient subspace. Experiments on BERT fine-tuning (GLUE) and GPT-2 pretraining (WikiText-103, enwik8, text8) show performance comparable to or better than standard SMoE. The method also works as a drop-in module for existing SMoE architectures.
In briefWhy long-context continual pre-training should not stop early: an empirical scaling law
SummaryThis paper studies how the benefits of long-context continual pre-training (LCCP) grow with the number of training tokens. The authors propose an empirical additive power law linking long-context performance to model size and short- and long-context token budgets, and introduce Constant-Last4, which averages the last four checkpoints of a constant-learning-rate run to monitor training stably. For a 1.22B-parameter model trained on about 100B LCCP tokens, gains show no clear saturation and persist after 128K context extension and supervised fine-tuning, extending to long-context understanding and agentic tasks.
In briefClosed-loop dual-reference guidance to control accents in cross-lingual voice cloning
SummaryZero-shot cross-lingual voice cloning often carries the reference speaker's accent into the target language. The paper proposes pairing the original reference with a native anchor built by voice conversion, using both to guide a shared acoustic trajectory through separate language-model states that are updated after each patch. On 20 held-out French speakers, the fixed setting reduces residual French LID mass by 73% versus direct cloning, with an ECAPA similarity cost of 0.042.
In briefB-CARE: Blind Ranking of Likely Errors in Stance Classification
SummaryThis paper introduces B-CARE, a framework that flags stance predictions most likely to be wrong so that limited human review can focus on them. A fixed LLM verifier scores each candidate stance label from only the target and text, without seeing the base model's prediction or confidence, and risk is computed by contrasting the strongest alternative label's evidence with the predicted label's evidence. Tested on four stance benchmarks and three base LLMs across 12 settings, it achieves a mean AUROC of 0.735 and captures 41.06% of errors in the top 20% of predictions. The method does not uniformly beat strong confidence baselines but offers a reusable complementary signal.
In briefShows that data schedules and learning rate schedules should be designed jointly during LLM mid-training.
SummaryThe paper studies data mixing schedules in LLM mid-training, which set how the share of pretraining and mid-training tokens changes over training. Under fixed compute and mid-training token budgets, it asks whether a complete switch to mid-training data minimizes final mid-training loss. The answer depends on the learning rate schedule: early mixing helps when the learning rate decays but reverses under a constant learning rate. Ablations suggest decaying the learning rate only after mid-training data enters is better, and a theoretical model is provided to explain this.
In briefBenchmarking and error analysis of deep search and deep research AI agents
SummaryThe paper notes that current evaluations mostly judge final outcomes and overlook errors in the intermediate steps of AI agents performing long, multi-step research tasks. It introduces a methodology and analytical framework for diagnosing these intermediate errors and a taxonomy dividing them into system-level and task-level errors. Through step-by-step analysis of execution traces and agent-environment interactions, the authors examine how these errors affect performance and provide design guidelines to improve robustness and reliability.
In briefDecodable Speaker Info Isn't Controllable: Local, Anisotropic Control in Continuous TTS
SummaryThis paper examines whether speaker information that is linearly decodable from continuous-latent text-to-speech (TTS) models can actually be used to control the output voice. In VoxCPM2, probes on the final residual acoustic language model layer reached 0.970 accuracy, yet aligned edits made the target speaker dominant in 0 of 60 directed speaker pairs. Pair-aligned updates at the earliest layer (L0) achieved target dominance in all 60 pairs, indicating that effective control requires geometric specificity and a route that carries the intended speaker condition into acoustic synthesis.
WhyShows linear decodability diverges from causal control, validated by position-matched interventions, independent speaker encoders, and double-blind listening tests.
In briefUncensored models still show script-dependent bias on Taiwan topics
SummaryThe paper tests whether abliterated (uncensored) versions of two open-weight 2-bit language models become free of political bias. Models were scored on statements about Taiwan, 1989 Tiananmen and control countries in Traditional Chinese, Simplified Chinese and English. The authors find that the writing script mattered far more than abliteration, which removed refusals but left the script gap intact. They conclude that uncensored is not unbiased.
In briefCoupling of readers and writers in transformer MLP layers
SummaryThis paper treats a transformer MLP layer as a key-value memory in which each hidden channel has a reader that decides when it responds and a writer that decides what it adds to the residual stream. By analyzing the reader and writer together through the exact Jacobian of the block, the authors assign each channel a signed pairing coordinate. They report that each block contains a small core of strongly coupled channels, mostly with opposed reader-writer directions, a pattern they say appears from small models trained from scratch to released checkpoints.
In briefChronoSynth: Constraint-aware question generation over temporal knowledge graphs
SummaryThis paper studies generating natural-language questions from temporal knowledge graphs while preserving the answer, time condition, reference event, and temporal order. The authors propose ChronoSynth, a memory-augmented framework that builds a Temporal Intent State, retrieves and adapts typed question patterns, and validates generated questions with feedback-guided revision. Experiments on CRONQUESTIONS and MultiTQ report state-of-the-art generation performance and ablations supporting its components.
In briefPolicy Central: Detecting Conflicts, Redundancy, and Similarity in Police Policy
SummaryThis paper presents Policy Central, a deployed system that helps police departments find whether sections of their written policies conflict, are redundant, are merely similar, or are different. It compares two architectures: Retrieve-Judge, a low-cost library-wide check, and Screen-Confirm, an exhaustive check of documents a reviewer selects. On a 302-pair benchmark from one department's roughly 84 policy documents, Screen-Confirm reached 0.795 aggregate recall given each pair and 0.885 on conflicts, but composed recall for finding the relevant pairs was only about 0.46 to 0.47.
In briefFrom Prompts to Posteriors: Can LLMs Do Bayesian Updating In Context?
SummaryThis paper evaluates whether large language models can update their beliefs in context according to Bayesian inference. It builds a framework combining probabilistic coherence tests, quantitative benchmarks, and analysis of how non-Bayesian updates affect downstream tasks. The authors find that LLMs generally violate Bayes' rule in context, but with modern reasoning abilities and code execution, they closely approximate analytically intractable posteriors.
In briefAgentMonopoly: how LLM agents behave and how persona instructions change them in a property-trading board game
SummaryThis paper uses AgentMonopoly, a game environment where AI agents buy, trade, and communicate over properties, to compare LLM agents from four model families at two sizes each. It finds that larger models do not consistently win more games, and that persona instructions change agents' wording and some actions but not always consistently across decisions. The authors argue persona following should be judged by actions and outcomes, not just language, and note results are limited to the tested checkpoints, prompts, and five seeds per configuration.
In briefHow LLM Agents Converge to Agreement Across Many Conversations
SummaryThis paper studies how two LLMs converge toward agreement during multi-turn conversations on subjective topics. Using a linear classifier on single-conversation activations, the authors find an internal agreement representation that does not reliably track external agreement at the individual conversation level. When internal representations are aggregated across conversations, external agreement can be predicted by a Markov process, and internal aggregated agreement relates to it through a power-law mapping. The authors also report that models express a stance change externally before their internal representations reflect it.
In briefOACap: Caption-guided active observation for detailed video description
SummaryThe paper proposes OACap, a caption-centered agent that requests targeted observations for unsupported or missing information, gathers evidence with spatiotemporal tools, and iteratively verifies and revises its captions. It also introduces FSCap-Bench, a benchmark with the Core-Support Factuality Score (CSFS) for claim precision and importance-aware fact coverage, and the Cross-Modal Semantic Round-Trip Score (CSRS) for semantic preservation through a video-caption round trip.
In briefSignals and Queries: a cheap first-stage filter for choosing which agent policies apply
SummaryThis paper introduces Signals and Queries (S&Q), a lightweight first-stage filter for user-facing LLM agents that decides which behavioral policies are relevant to each turn. Offline, an LLM generates activating examples and hard negatives for each policy, and an independent logistic classifier on their embeddings predicts relevance without per-turn LLM calls. Evaluated on 9,144 LLM-judge labels across 254 conversation prefixes and 36 airline policies, S&Q removes at least 30% of candidates while preserving high recall and achieves higher precision than top-K and cosine-threshold retrieval, consistently across two embedding providers.
In briefOpenMAP-HIE: Unsupervised Neonatal Brain Injury Lesion Segmentation Using Pathology-Informed Diffusion Models
SummaryThis paper proposes OpenMAP-HIE, a score-based diffusion model that learns the distribution of healthy neonatal ADC images and locates hypoxic-ischemic encephalopathy lesions from discrepancies between conditional and unconditional noise predictions. It needs no voxel-wise lesion labels and keeps only discrepancies matching the expected ADC reduction in acute HIE. Evaluated on the BONBID-HIE dataset, it achieved the best overlap and surface-distance results among unsupervised methods. Adding population-derived ADC Z-map refinement raised the Dice score to 0.602, though small and low-contrast lesions remain difficult.
Pengyu Kan (Johns Hopkins University), Craig Jones (Johns Hopkins University, University of Alberta) et al.
In briefBrainCodes: A Lightweight EEG Foundation Model Built on Discrete Neural Codes
SummaryThis paper introduces BrainCodes, a compact EEG foundation model pretrained with self-supervised learning on over 800 hours of recordings from more than 1,000 participants. It combines a quantization-based neural tokenizer with a spatiotemporal encoder that handles arbitrary electrode configurations. The model matches or surpasses prior state-of-the-art results on tasks such as sleep staging, attention assessment, and cognitive load estimation. It transfers to low-resource wearable devices while using up to 97% fewer trainable parameters than comparable architectures.
Kleanthis Avramidis, Ivan Tashev, Dimitra Emmanouilidou
In briefOSynth: A synthetic dataset for testing how well systems understand meetings using background context.
SummaryThis paper introduces OSynth, a large synthetic dataset simulating organizations, with 570 hours of meeting audio, related artifacts, and a latent ground-truth state. In pairwise comparisons, human raters identified OSynth meetings versus real transcripts with only 56% accuracy, using 150-300 word excerpts. The authors report that adding context significantly improves state-of-the-art LLMs on several entity-sensitive dimensions, suggesting OSynth can support better context-aware meeting understanding systems.
Ian Porada, Suraj Tripathi, Chaitanya Animesh et al.
In briefAsking what a director sees can change which block vision-language models choose
SummaryThe paper studies whether vision-language models pass theory-of-mind tests, using a rendered Director Task with a partition hiding the largest block. Asking the model first which blocks the director can see sharply raises the rate at which it selects the hidden block, even when its reports are mostly correct. The authors argue that evaluations eliciting a perspective report before scoring an action need a baseline without that question, and they provide paired prompts differing only in that question.
In briefSpatialBrain: Spatial Intelligence That Grows From Experience
SummaryThis paper presents SpatialBrain, a framework that accumulates experience about how to make spatial reasoning decisions, such as when to observe directly versus reconstruct, measure, or change reference frames, while keeping the vision-language backbone fixed. It combines a 3D memory that fuses current-scene observations, a library of reusable spatial tool procedures, and a cognition module linking task conditions to memory and skill use. Across five spatial-reasoning benchmarks, it outperforms direct answers from the same backbone on four and ties on the fifth.
In briefDISRL: A Reinforcement Learning Method for Tracking Communities in Changing Networks
SummaryThis paper studies dynamic local community detection, which tracks the community of designated seed nodes across evolving graph snapshots. The authors propose DISRL, which includes a module that generates seed-specific guidance from sparse known communities and a dual-scale exploration module that separates short-term adaptation within time windows from long-term parameter consolidation across windows. They also build a dynamic graph benchmark with diverse community evolution scenarios. The paper reports that DISRL consistently outperforms state-of-the-art baselines under various dynamic settings.
In briefYingLan: A multimodal ocean foundation model learning unified representations from varied ocean data
SummaryThis paper presents YingLan, a multimodal ocean foundation model that learns unified representations from heterogeneous data such as satellite imagery, gridded oceanographic fields, and buoy observations. It combines physics-attribute-conditioned dynamic encoding, Spatiotemporal Increment Coupling (STIC), and pretraining that reconstructs masked observations and models ocean-state evolution. In a seven-day sea surface salinity forecasting task in Shenzhen coastal waters, it achieved an average RMSE of 0.687508 PSU, with relative error reductions versus LangYa, XiHe, and GloNet2. The authors note that broader cross-task and cross-regional evaluations are still needed.
In briefInterpretable genetic-programming cut selection for mixed-integer linear programming
SummaryThis paper proposes GPSCS, a framework that uses genetic programming to evolve an interpretable symbolic function for ranking candidate cutting planes, and a statistics-based mechanism to adaptively choose how many top cuts to add. Tested in the SCIP solver, evaluation instances are on average 41 times harder than training instances. The method improves average solving time over SCIP's default cut selection and achieves about a two-fold speedup over a neural-network approach and a no-cut baseline.
In briefDifferentially private synthetic data using generalized moment matching
SummaryThis paper studies generating (ε,δ)-differentially private synthetic data from arbitrary databases of n records in [-1,1]^d, under a specified integral probability metric (IPM). The authors introduce generalized moment matching, which privately estimates moments, fits a probability measure, and samples synthetic records. The framework extends Chebyshev moment matching and is applied to maximum mean discrepancy with Matern and Laplace kernels, a two-layer ReLU neural tangent kernel, sliced Wasserstein-1, and IPMs from norm-bounded ReLU networks. With enough synthetic records, the rates are minimax optimal up to logarithmic factors for fixed dimension.</summary>
In briefLLM-Based Exogenous World Modeling for Dynamic Scheduling
SummaryThis paper studies dynamic scheduling, where new jobs arrive over time and must be dispatched while anticipating future arrivals. The authors use a pretrained large language model to forecast uncertain future job arrivals within a structured world model. They propose Afterstate World-Model Planning, which combines a belief-conditioned dispatching strategy with short scenario-based rollouts. Across various arrival distributions, the method reduces weighted tardiness relative to the best of eight classical priority-dispatching rules on dynamic and flexible job-shop benchmarks.
In briefHindcasting measured wave spectra with data-driven and physics-guided models
SummaryThis paper tests a recurrent model that starts from an hourly buoy wave spectrum and uses local wind, fetch and depth to produce short-range hindcasts. Measurements from three Baltic Sea stations were used for training, validation and testing across 2017 to 2022, with ERA5 winds. A local relaxation constraint gave modest, objective-dependent improvements in wave height and spectral error. Operational forecasting and transfer to independent stations remain to be evaluated.
In briefExpEvolve: Structuring and reusing experience for open-ended discovery
SummaryThis paper proposes ExpEvolve, a method that improves LLM-driven evolutionary program search by extracting high-level "idea factors" that explain score differences between programs. These factors are used to guide exploitation of past experience and exploration when progress plateaus. Across 11 open-ended mathematical tasks, ExpEvolve improves final performance and sample efficiency by up to 5.0x. The gains also transfer to 4 other LLMs and 6 machine learning tasks from MLS-Bench.
In briefFixing Multimodal Model Edits That Rely on Background Clues
SummaryThis paper shows that multimodal language model edits can produce correct answers while depending on irrelevant background context instead of the intended visual evidence. The authors build WaterBird-Edit, an analysis set that separates bird habitat labels from background to test this. They propose VICE, a retrieval-based method that stores edits externally and uses visual in-context demonstrations, which performs strongly on three benchmarks and reduces sensitivity to background shifts.
In briefCompressing onboard language models for space robots while preserving behavior
SummaryThis paper studies how to compress language models to low-bit sizes so space robots can run them onboard to interpret operator commands and select verified skills. The authors propose QSelect, which combines layers from AWQ, GPTQ, and AutoRound quantized checkpoints into one static model, scoring each substitution by its final output distribution against the full-precision reference. Experiments use general-domain benchmarks on a ground GPU, and the composed models achieve lower perplexity than any single-quantizer model. Space-domain tasks and flight hardware evaluation are left for future work.
In briefHow close to Nash? A certificate for solutions in general-sum games
SummaryThe paper proposes an online certificate that uses the regret accumulators already maintained during solver updates to bound the NashGap of a returned strategy profile, without computing best responses against the average strategy. The certificate is shown to remain sound under alternating updates and in games with three or more players. Its effectiveness is evaluated on a biased Leduc benchmark using a portfolio of training-free regret-matching solvers.
In briefLatent space configuration for faster neural network embedding search and dimensionality reduction
SummaryThis paper proposes search algorithms for structured latent spaces of neural networks that speed up retrieving the top K most similar items by searching only a small candidate subset. The authors finetune ArcFace-pretrained IResNet50 models with the Latent Space Configuration method to obtain embeddings with desired geometric properties. Compared with the baseline exhaustive search, the authors report up to 3 times speed-up on GPU and 40% on CPU. Reducing the latent space dimensionality further accelerates search and can even improve rank-N metrics.
In briefSame Odds, Different Advice: How Gain or Loss Wording Shifts LLM Guidance on High-Stakes Decisions
SummaryThis paper examines whether large language models give different high-stakes advice when the same probability is framed as a gain or as a loss. The authors built FrameBench-HPS, a benchmark of 2,880 prompts from 240 scenarios in healthcare, finance, education/career, and legal/consumer rights, and tested seven LLMs. Gain-loss framing changed the recommendation category in 32.4% of complete groups, mostly through uncertainty rather than outright reversals, and probability canonicalization substantially reduced this sensitivity.<br>
In briefGiving Credit Where It's Due: Contrastive Search for Multi-Agent Teams
SummaryThe paper proposes CATS, a method that searches for compact teams of specialized LLM agents within a fixed supergraph. An LLM attributer compares execution traces across candidate topologies, assigns credit to individual agents, and uses these credits to update edges and prune weak agents and edges. On BigCodeBench, CATS improves mean accuracy by 5.3 points over score-only heuristic search. In a HotpotQA prompt-poisoning stress test, CATS excluded all adversarially prompted agents in all three runs.
In briefApproximate Solomonoff Induction: How Robust Learned Universal Predictors Are to Approximation Error
SummaryThis paper examines whether prediction performance stays robust when learned predictors only approximate the optimal Bayesian mixture, which corresponds to Solomonoff universal induction. It proves tight bounds showing that approximation can substantially degrade performance in environments with small prior probability, while the excess error averaged over the prior is only O(η). The authors also give a distribution-shift result and extend the analysis to finite-sampling settings relevant to neural networks and Transformers.
In briefVision and knowledge-based language models to help beekeepers inspect hives
SummaryThis paper presents an edge-deployable pipeline that turns photographs of hive frames into interpretable, action-oriented inspection reports. An object detector identifies comb-cell classes, the detections are converted into structured facts and constrained text, and a grounded knowledge base provides interpretations and recommendations. The system is built on the open FAIRHiveFrames-1K dataset, and the authors release FAIRHiveText-1K, which pairs those facts with text.
In briefFlat-RLVR: Sharpness-aware reinforcement learning to keep LLM reasoning diverse
SummaryThe paper notes that reinforcement learning with verifiable rewards (RLVR) can improve LLM reasoning but tends to sharpen the policy distribution, suppressing rare yet correct trajectories and reducing sample diversity. The authors propose Flat-RLVR, which uses sharpness-aware minimization during post-training to seek flat reward regions, combined with GRPO variants. They report improved sample diversity while maintaining single-attempt accuracy, and theoretically link flat-reward optima to reduced distribution sharpening.
In briefBayesian Evolutionary Optimization for Designing and Tuning Multi-Agent Systems
SummaryThis paper proposes a Bayesian evolutionary optimization framework that automatically discovers task-specific multi-agent system topologies. It combines evolutionary operators that generate diverse workflow graphs with a graph-aware Gaussian Process and Expected Improvement to focus costly evaluations on promising candidates. On BigCodeBench, evolved architectures achieved the highest overall performance among evaluated agent systems, and on temporal tabular datasets the optimized preprocessing improved over default TabPFN on several tasks.
